PDF OCR Web Service

A lightweight web application for OCR (Optical Character Recognition) of PDF documents and images powered by Qwen2.5‑VL‑7B‑Instruct (vision‑language) and Qwen2.5‑7B‑Instruct (text‑only) models.

---

Key Features

• Upload PDFs or standalone images.
• Render any PDF page at 150 DPI and run OCR on the current page or the whole document.
• On‑the‑fly text clean‑up and Polish summaries generated with an LLM.
• Responsive UI with progress bar and cancellation.

---

⚙️ System Requirements

Item	Minimum	Recommended
OS	Linux (Ubuntu/Debian 12 tested)	Any modern Linux
GPU	NVIDIA 24 GB VRAM (RTX 3090/4090)	≥ 48 GB VRAM for long docs
CUDA	12.1 +
Python	3.10 or 3.11

Why CUDA 12.1 +? flash_attention_2 (used by Qwen2.5‑VL for faster inference) requires a recent CUDA toolkit

---

Python Dependencies

Library	Notes
PyTorch ( GPU build)	torch ≥ 2.2 compiled for CU 12.1
flash‑attn (optional → speed)	pip install flash-attn --no-build-isolation
transformers (dev)	Install from source to get qwen2_5_vl classes
accelerate	Device placement & 4‑bit loading
qwen‑vl‑utils[decord]==0.0.8	Helper for process_vision_info
bitsandbytes	4‑bit quantisation support
flask, pillow, pymupdf	Web server, image IO, PDF rendering

A ready‑to‑use Conda environment file is provided below (environment.yml).

---

Quick Installation

# 1. Install Miniconda (skip if already installed)
wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
bash Miniconda3-latest-Linux-x86_64.sh -b -p $HOME/miniconda3
eval "$($HOME/miniconda3/bin/conda shell.bash hook)"

# 2. Clone the repo and create the env
git clone https://github.com/yourusername/PdfToOcr.git
cd PdfToOcr
conda env create -f environment.yml
conda activate ocr

Manual package install (alternative)

conda create -n ocr python=3.10 -y
conda activate ocr

# GPU PyTorch + FlashAttention build
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
pip install flash-attn --no-build-isolation   # optional but recommended

# Latest Transformers w/ Qwen2.5‑VL
pip install git+https://github.com/huggingface/transformers accelerate

# Rest of the stack
pip install qwen-vl-utils[decord]==0.0.8 bitsandbytes flask pillow pymupdf

---

Running

conda activate ocr
python pdf_server.py

Navigate to http://localhost:5000.

For production, enable the provided systemd unit (ocr.service).

---

Data Paths

Uploads + page PNGs are stored in ./uploads/.
Make sure this directory exists and is writable.

---

🔍 environment.yml

name: ocr
channels:
  - conda-forge
  - nvidia
  - pytorch
dependencies:
  - python=3.10
  - pip
  - pip:
      # GPU PyTorch CU12.1 wheels
      - torch==2.2.*+cu121 torchvision==0.17.* torchaudio==2.2.* --index-url https://download.pytorch.org/whl/cu121
      # Speed‑ups
      - flash-attn
      - bitsandbytes
      # HF ecosystem – dev transformers for Qwen2.5‑VL
      - git+https://github.com/huggingface/transformers
      - accelerate
      - qwen-vl-utils[decord]==0.0.8
      # App stack
      - flask
      - pillow
      - pymupdf

Create the env with:

conda env create -f environment.yml

---

References

• Qwen2.5‑VL model card 
• Transformers docs for Qwen2.5‑VL

---

Happy scanning! 📄✨